This release targets users who need a local, high-context, multimodal Qwen3.8 variant that prioritizes direct answers and throughput over refusal-based safety preambles. The package combines per-quant GGUF builds, a BF16 vision projector for image/video inputs, and an optional FastMTP draft sidecar that speeds speculative MTP decoding while leaving the verified target model unchanged.
What Sets It Apart
- Aggressive uncensoring profile: configured to produce direct answers with minimal preamble and no built-in refusal behavior, useful where compliance prompts would otherwise block useful completions.
- HauhauCS FastMTP sidecar: a 32K draft profile that can accelerate document and reasoning throughput (benchmarked up to ~3.02x document TG and ~1.93x reasoning TG versus MTP-disabled runs), while the full target still verifies each token.
- K_P quant family and variants: per-quant "Perfect" (K_P) quantizations selectively preserve critical tensors to raise quality by ~1–2 quant levels at modest size overhead; many quant sizes offered to fit different GPU memory targets.
- Native long-context and multimodal support: the underlying Qwen3.8 architecture supports a very large native context (262,144 tokens) and includes a vision projector for BF16 image/video inputs.
Who It's For and Trade-offs
Great fit if you need a local multimodal Qwen3.8 replica with long-context, configurable quant/size trade-offs, and optional speculative MTP acceleration for higher serving throughput. It is practical for benchmarking, research, and inference workloads where direct answers and throughput matter more than built-in refusal constraints. Look elsewhere if you require a safety-first/default-refusal model, strict moderation by default, or if you cannot run patched/runtime setups (FastMTP requires a patched llama.cpp runtime to use the sidecar). Also plan VRAM and context sizing carefully: large native context and K/V precision raise memory needs.
